Superpixel-based automatic segmentation of villi in confocal endomicroscopy

Confocal Laser Endomicroscopy (CLE) is a technique permitting on-site microscopy of the gastrointestinal mucosa after the application of a fluorescent agent, allowing the evaluation of mucosa alterations. These are used as features by skilled technicians to stage the severity of multiple diseases, celiac disease or irritable bowel syndrome among the others. We present an automatic method for villi detection from confocal endoscopy images, whose appearance changes with mucosal alterations. Superpixel segmentation, a well-known technique originating from computer vision, is used to identify and cluster together pixels belonging to uniform regions. Each image in the dataset is analyzed in a multiscale fashion (scale 1, 0.5 and 0.25). From each superpixel, 37 features are extracted at multiple image scales. Each superpixel is classified using a random forest, and a post-processing step is performed to refine the final output. Results in the test set (70 images, 30870 superpixels) show 85.87% accuracy, 92.88% sensitivity, 76.99% specificity in the superpixel space, and 86.36% of accuracy and 87.44% Dice score in the pixel domain.